I think my favorite part of buildzoid is "let's try to do this more simplified and concise", then proceeds to say "let's start with opening three engineering data sheets"
TL:DR Due to it being faster for Memory to use a different bank it is better (in terms of bandwidth) to have more banks Think, its always faster to switch to your secondary! x4 is for servers to enable high density sticks x8/x16 is consumer grade of which: x8 is preferred as it has 4 banks. x16 is a "space saver" and as a result only has 2 banks. So if possible you want 1R8 or 2R8 (again consumer grade) Also not all ram manufacturers list specs nicely or can use different configurations based on their revision number!
This is pretty great summary. I noticed even on the original Linus video a lot of people confused number of bank groups with memory ranks, which obviously are not the same thing. Now if only it was easy to get all the specs for retail memory.
Linus was confused, they did not seem to know what the problem was. The real problem is going to be that each week the ram in a laptop may be a different brand. PS to be fair it's above my head, ill take bulldozes word for what is faster. edit for the laptops it comes down to, was it the old problem of tricks being played to make the Nvidia laptop faster in benchmarks or was it random luck from that week of production that the Nvidia laptop had the faster ram? Suspect it's old tricks as Nvidia always seems to win, but may be luck. 20% faster with the 'good' ram is a lot!
@@defeqel6537 Im not up on laptops but from what iv seen it's normal to change as ram comes in. Unless there sourcing from some one like Samsung or crucial it may not even be up to the laptop OEM what ram they get, lots of ram brands are pot luck on what ram brand is used on the sticks. Iv even seen people say some laptops change from dual to single sticks and back during there life cycle. Never mind during a parts shortage where it may not even be an option to stick to a single ram kind. (and that ignores that most OEM's will at least dual source parts)
@@ActuallyHardcoreOverclocking well, whatever you want to call it, I appreciate that you've put out a video explaining memory ranks that I can point people to without worrying that they won't understand.
Every other youtuber: "Oh man...I hope I can manage to squeeze at least 10 minutes out of the footage" Buildzoid: "Oh man...I hope I can manage to keep focused and don't babble for 1.30 hours for this footage"
I think the biggest bullet point about RAM in this context is that most "normies" don't know that the bus width of DDR is 64-bit. This is then split into the above. 64-bit split into x4 is 16 modules, x8 is split into 8 modules, and x16 is split into 4 modules. EDIT: First of all, thank you for the heart Buildzoid. Since this is going to likely be a featured comment for a while, I would like to point out that what is being talked about in this video has nothing to do with memory ranks. The LTT video wasn't about ranks. Buildzoid's response to that video wasn't about ranks. This is about the SDRAM ICs on the DIMM, and the available 64-bit bus being split between all the modules on the DIMMs; see my original post for that summary.
@@ActuallyHardcoreOverclocking for now at least. There's alot of people working to make RAM persistent. Game changer, no longer having to rewrite to each address row every y seconds depending the weakest capacitance in the row.
TL;DR: 64-bit RAM addresses are broken into x16, x8, or x4 bit banks, where each bank has a cost to access, write, read, and maintain data. The cost of performing a subsequent operation in the same bank is more than doing it in a different bank, thus the more banks, the better. As a result, x8 is better than x16 RAM configurations. NL;PR: Random-access memory performance is directly proportional to the available slots to operate at once: dual-channel, x8 bit banks, dual-rank (2R) is the best consumer option there is so far; and to how often those slots can be operated on: higher frequency (>hz), lower latency class (
Found it weird, that you had to make a new video explaining. Really think you did an awesome job in the other video. I'm really unfamiliar with the topic and understood every thing you were trying to explain, clearly.
and I would say it's a great thing. That's how everyone becomes more educated on the topic i.e. how manufacturers are screwing their customers without letting them know explicitly
Thank you for a really useful video, any chance you could do a follow up on how i could identify on desktop ram if a kit is x8 or x16 as i cant see an easy way to identify this before purchase
@@claritoresdiano1021 Can you help me out pls ? I have soldered memory of 4gb 2133mhz, 1Rx16 on my laptop. Should I add 4gb of x8 memory to it or 8gb of x16? (I cannot find 4gb of x16 in market)
@@akhilsingh9884 Open command prompt (Run As Administrator) then type: wmic memphysical get MaxCapacity, MemoryDevices to see the Maximum Capacity for each slot and the number of slots you have on your motherboard. You could afford 8Gb x16 just go on that you get more headroom. (4GB onboard) + (8GB Dedicated slot) = 10GB. Don't bother about 4Gb x8 because your system can not maximize it.
The big question is how do you know in advance whether its x8 or x16. I am looking at desktop RAMs and none of the big manfacturers (Corsair, G.Skill, Crucial) mention this on their website
See BuildZoid's other video. tl;dr: E-bay Corsair memory kits and look at the version in the listing's picture. You can look up what version corresponds to what memory module. The same exact SKU/product varies widely, so that's the only way. Alternately, you can try to find OEM or server parts, since they list plenty of extra details, but they're probably either going to be slow or expensive.
Each pin on a memory IC has a tiny bit of capacitance (a parasitic property of the FET gates) and every time you add another memory chip to the bus you increase the total capacitance on the memory bus lines. The extra capacitance causes additional electrical "load" on the line, leading to slower transition times between defined voltage states (0 and 1). The longer the transition time, the less time the bus lines spend at a defined "1" or "0" state, making it more prone to errors. This is usually measured as an eye diagram. If you add a ton of memory ICs, you eventually end up way out of spec and the system can't function reliably. In an unregistered DIMM (UDIMM) you've got a bunch of memory chips and they all share address lines, command lines, and (to some extent) data lines. All of these ICs have some parasitic capacitance that contributes to line load. Registered DIMMs (RDIMMs) get around this by adding a buffer chip (a "register") between the memory ICs and the memory bus. Instead of connecting all the command and address lines from all of the chips to the memory bus, they instead connect them to a single buffer chip, and then connect that to the memory bus. The data lines remain connected directly from the memory ICs to the memory bus. The buffer distributes the address and command signals to the chips, which is easier for the buffer chip than the memory controller because it is physically close to the memory ICs, meaning there's less impedance in the signal path, and it can also have separate IO lines for groups of memory ICs, decreasing the load on each driver in the buffer. This approach heavily reduces the amount of electrical load on the command and address lines, allowing you to add more memory ICs in total, for better overall memory density. There's a latency penalty due to the added buffer, and the modules use a little bit more power too. Load reduced DIMMs (LDRIMMs) go one step further and buffer the data lines too. This is more expensive since the buffer needs to handle many more signals, but it vastly reduces the bus load for all of the high speed signals, including data. This allows you to have far more memory ICs per module (technically it's more ranks, but explaining ranks is... ugh) and therefore huge memory densities on the standard DDRx bus. The penalty, again, is latency and power. Plus extra cost, of course, but that's usually not as much of a consideration when you're looking at 128GB DIMMs. A rule of thumb is that UDIMMs use low density 8/16-bit chips, RDIMMs support high density 4-bit chips but don't support a lot of ranks because the data lines aren't buffered, and LRDIMMs support 4-bit chips AND high ranks (up to 8 iirc) because they buffer the data lines as well as the command & address lines.
(aside: there's also FBDIMMs but they're super frickin' weird and afaik nobody uses them any more. they use a serial bus to lower the number of lines that need to be driven, then expand that back to a parallel DDR bus on the DIMM using something similar to an LRDIMM register buffer but waaaaay more complicated. a lot of that tech ended up being vapourware.)
If there is any interest, could you do a partpicker walkthrough of available 2x32GB kits? I was looking through your previous videos and didn't find anything like this specifically.
This video isn't really a simplified version of the other, it's a complement. The video explains why the realities of manufacturing dictate why x16s have a performance penalty while the other has discussion on the qualities of different dies and whether this information is made available to the consumer.
Thanks. That made sense. It could also be helpful if you started with a photo of an RAM entire module, and then break that down into ranks, and then into the individual chips, and then into the banks in the chips. Perhaps with examples of 1Rx8 vs 2Rx8 (dual rank ryzen memory debacle) vs 1Rx16 (stupid laptop manufacturer debacle)
Huh, I thought the original video was perfectly understandable. How I interpreted it: if you want a faster HDD array, add drives so you can have more seeks going on at once; if you want a faster RAM array, add bank groups so you can have more selects going on at once...
sort of, yes, because then the memory controller would have 2 ranks with 2 bank groups each to use instead of 1 rank with 2 bank groups. But for the applications where this could matter (laptops etc) you won't have the opportunity to do that because you're limited to 2 SODIMMs anyway and they'll be single rank. For desktops you'd much rather have 2 ranks with 4 bank groups each if they actually made x16 chips with 4 bank groups instead of 2, then they could make dual rank SODIMMs (same number of chips as single rank with x8 chips) with 4 bank groups which would be pretty awesome for laptops
So glad you made this video. 🙏 Can x4 theoretically give a performance boost if the density is kept low? I have been trying to understand the RAM options for the Mac Pro 7,1 for a long time! Because nerdy.
x4 = 16 chip x8 = 8 chip x16 = 4 chip Theoretically is yes but the fact is impossible. SODIMM is small no headroom for 16 chips (x4), that's why just x8 or x16
I do have x4 laptop sticks because the old PoS Core 2 CPU in an old laptop I have can't handle x8 4GB DIMMs, but is fine with x4 4GB DIMMs, so it was the only way to get 8GB in that machine. The only reason this travesty even exists is because some old Macs had the same issue and of course HW manufactureres will make any garbage for Macs (it's actually sold as "Mac memory").
I don't think the problem is tccd. I think the issue is the fewer number of banks increases the probability of needing to precharge the current row and activate a different one. The tccd seems to me like it would be way easier for the controller to schedule around.
Can you help me out pls ? I have soldered memory of 4gb 2133mhz, 1Rx16 on my laptop. Should I add 4gb of x8 memory to it or 8gb of x16? (I cannot find 4gb of x16 in market)
@@ActuallyHardcoreOverclocking ah! we went to the same school.. although we moved on and now your a big guy on youtube and im a person watching your videos.
Hello, Have a look at the EDC bug on the 3000 ryzen. I have my 3900x boosting to 4.7ghz on half of the cores crazy stuff. Just use pbo, put EDC 1 clock to +500mhz escalar to 10* and then I did a bit of undervolt -0375 and push the bus clock to a point before 101,00.
What is this autogenerated Vietnamese subtitles? Why not english for non native speakers? it is night and day difference when there is proper transcript and I put on 0.75 speed and replay as needed.
Ok so if i get this correctly, if some company decides to make x16 chips with 4 banks, they'd be (approx) the same as x8's right? (because they have the same amount of bank groups) And is it possible to have more than 4 bank groups or is that outside of the spec? (idk if this got answered in the long vid, that one did go a lil over my head)
Although I am late with this comment, I would like to thank you for the in-depth explanation. It makes things clearer, especially that a branded memory stick is almost always x8 but non-branded may very well be x16. But I do want to ask how you know what type you have in your laptop. Do you really have to open it up and read the sticker? Is there no software like CPU-Z or HWInfo64 that can tell you? I may be a n00b, but I can't find this info from these two programs.
Can you help me out pls ? I have soldered memory of 4gb 2133mhz, 1Rx16 on my laptop. Should I add 4gb of x8 memory to it or 8gb of x16? (I cannot find 4gb of x16 in market)
Akhil Singh in my case I dont see too much of a difference for everyday things that I do, mainly browse the internet and the only game I play is MW2 which is an old game which is not demanding, its always nice to have more RAM for when you do need it, my RAM was not soldier so that was a plus. I just match the one i ready had.In my case Ive read that AMD Ryzen CPUs are more sensitive to the right RAM.I guess with Intel you can mix and match different RAM modules. Having said that tho if you mix x8 with x16 you wont see the speed inmprovements of x8, Im not a expert but I know for my laptop I got the best RAM I could get, another thing my RYZEN cpu cannot use 3200mhz, I was gonna buy that speed because it was the fastest but I read that 2600mhz was the maximum, so if i had bought that speed I would have just wasted money, I would advice watching more RAM videos here on TH-cam, about your particular laptop and CPU, theres lots of giod videos here
TH-cam: This guy has a terrible accent don't you think? Caption algorithm: Sure, somewhat asian style, I heard Micron and Samsung. TH-cam: Let's assume he's vietnamese then. Caption algorithm: Agreed. Starting auto-translation... "that's what going on with a, with-aaa bus width... ghi chép tât phen nuôi work best" Matching voice 99%.
Manufacturers can solve this problem with 3D layered chips or stacking more die in a single package. I.E. High capacity are typically stacked in an 8 die per package format. The would need to use a higher density BGA.
One rank is 64 bits of memory interface. So x4, x8 and x16 only relates to rank in the sense that it dictates how many physical modules are needed to reach the 64 bits needed per rank.
Say you have a few filing cabinets at your office, but if you open them too quickly it jams, so you have to do it slowly. Now lets say that Corporate has come down and put a thing that auto-closes them after every page you take it/put in. It takes 4 seconds to open a cabinet, and 2 seconds for the auto-closer to work. So, if you only have a single cabinet, you would have to wait 6 seconds between each page. However, if you had two cabinets, then you could try to spread the document between them, and only have to wait 4 seconds. Except, sometimes you end up needing two pages from the same cabinet. Well, with four cabinets, that's much less likely to happen. So, it's faster. That's an analogy for what's happening. x16 is like having two larger cabinets instead of 4. They can each hold more pages, but you're much more likely to have to wait the full 6 seconds. Incidentally, in this analogy, the cabinets are RAM. Your desk is cache, and corporate's main document repository across the country where documents have to be overnighted is your SSD.
Your comment about servers using x4 is a bit inconsistent. An x4 chip would have, going by the 4Gbit x4 example, the same capacity as the 2Gbit x8. You can see this on the datasheets: 262144 x 128 x 32 = 1,073,741,824 bits, is the same as 131072 x 128 x 64. Not sure really why you'd use x4, maybe easier PCB layout for lower speed controllers. Or maybe they're cheaper to make for non-dense applications.
x4 chips require 16 chips per rank which is hard to drive so they only get used on RDIMM. RDIMMs are for severs because consumer chips don't even support those.
You have to have 64-bits worth of memory ICs in a rank. If you use 8-bit ICs, you can only have 8 of them comprising a single rank. That limits the total per-rank density to the maximum capacity of available memory ICs multiplied by 8. If you use 4-bit ICs, you can double that and get twice the density per rank. The problem is that all of those address and command lines are shared, so the load on the bus (largely due to parasitic capacitance on the memory IOs) is doubled when you do that. That's why you only see them in buffered memory (i.e. registered DIMMs) and not in consumer memory.
@@ActuallyHardcoreOverclocking Right - I get that. My comment was more around how they'd put more chips on an RDIMM to get higher memory capacity, but this wouldn't work like that. Each chip has the same capacity, just differing geometry. The only reason to do this that I can see is if you wanted a 32-bit bus instead of e.g. a 64-bit bus, but the same capacity. Maybe for an embedded processor that didn't have the pins for 64 bits.
@@ActuallyHardcoreOverclocking I think you're missing out a bit here,my laptop has x4 memory 3200mhz 8gb each,if I go by your facts this ram shouldn't exist.
Thanx man! It was explanation on ranks and banks, I was just reading about ram and came across this term "absolute latency" calculated from ram speed (MHz) and its CAS latency (ns) on online ram latency calculators... and i was wondering is that a good enough indication of ram speed overall ?
The equation is * = . Where the chip width is x4, x8, x16, and the bus width is always 64 per stick of ram. So it's 64/ = . That means: x4 = 16 chips x8 = 8 chips x16 = 4 chips You can't physically fit that many chips on a SODIMM.
You said that x4 are density optimised and x16 are for small PCBs. From my understanding the goal in both cases is to cramp es much bit cells in a given space. Why is one ideal for server and the other ideal for Laptops?
If people didn't understand your first video....... well, you see where this is going. It's like teaching algebra to dogs, cute & entertaining but ultimately futile.
Amazing video, thank you Buildzoid! This makes me wonder about CPU Cache differences between L1 L2 L3 or even just more modern CPUs and older CPUs. It also makes me think about Nvidia’s various Graphics Card SKUs and how they have all these different memory bus widths, but it’s a set amount per 2 chips and that dictates capacity with the exceptions of higher capacity chips as found on the 3060 (non-Ti) and 3090.
@@ActuallyHardcoreOverclocking Oh I assumed it was double the density kind of like the RTX 2080 Ti 11GB to RTX TITAN being +1 chip to get 24GB. I didn’t think the 3090 had space on the PCB... hmm double sided i guess? Anyways thanks for pointing that out.
I think you could make this video much simpler and much better. Basically you just waved your mouse around and didn't explain the diagrams at all. Maybe you expected me to pause it and find the differences myself. Then you say they're the same chip so I guess the difference is just in the start of diagram, I don't know. You didn't say look over here, there's X amount of banks or X amounts of data lines or whatever. You just talked a bunch just to say they're the same chip that can be configured differently. I'm not sure if you explained why the timing is different based on configuration used. It seems like you just said one configuration needs to wait more clock ticks to access same block or bank or whatever that you didn't explain. It would of been nice to know what the block sizes are and so I get an idea how often it will try to read from same block or sector or whatever you call it. Part of me wants to be nice and try to help you and part of me is like omg that's the worst explanation ever so you realize how bad it was so I guess I'll just say both. I suggest just draw 3 squares, put 4x 8x and 16x then inside draw the blocks all the same size, tell us size, maybe draw a controller block and say this addresses them different. I rather a rough drawing that is explained then a complex one that is not.
I bought a sodimm from adat 8g 3200mhz it says 8Gx4 on the ram sticker and the chips are only on one side of the ram if I have a xr ram is it much better than x8?
If you ever find a compact notebook with the RAM soldered down to the motherboard it may be x16 simply to make room. Linus did a video on x8 and x16 RAM in identical computers. one came with x8 and one came with x16. x8 is seemingly better for performance somehow. It really is ironic but you would think the higher the number the better it surely must be. If you have one of those people who keeps saying more is better and doesn't know much about computers just send him on a wild goose chase looking for x32 RAM (which does not exist hehehe!!). Tell him not to come back until he finds it
...As a follow up to my previous comment, my main concern is that you'll start spending more time explaining basic s*it and less time being cryptically hardcore...which is why i tune in. Just keep being you and don't worry if the norms are keeping up. Google is a thing and learning is a skill that needs daily practice.
I think my favorite part of buildzoid is "let's try to do this more simplified and concise", then proceeds to say "let's start with opening three engineering data sheets"
Right jus tell me how to tell and Wich one definitely is the best that's all we need to know
TL:DR
Due to it being faster for Memory to use a different bank it is better (in terms of bandwidth) to have more banks
Think, its always faster to switch to your secondary!
x4 is for servers to enable high density sticks
x8/x16 is consumer grade of which:
x8 is preferred as it has 4 banks.
x16 is a "space saver" and as a result only has 2 banks.
So if possible you want 1R8 or 2R8 (again consumer grade)
Also not all ram manufacturers list specs nicely or can use different configurations based on their revision number!
*saw 11 minutes video in the thumbnail
"Wait is this Buildzoid?" :p
This is pretty great summary. I noticed even on the original Linus video a lot of people confused number of bank groups with memory ranks, which obviously are not the same thing.
Now if only it was easy to get all the specs for retail memory.
Linus was confused, they did not seem to know what the problem was.
The real problem is going to be that each week the ram in a laptop may be a different brand.
PS to be fair it's above my head, ill take bulldozes word for what is faster.
edit for the laptops it comes down to, was it the old problem of tricks being played to make the Nvidia laptop faster in benchmarks or was it random luck from that week of production that the Nvidia laptop had the faster ram?
Suspect it's old tricks as Nvidia always seems to win, but may be luck.
20% faster with the 'good' ram is a lot!
@@liaminwales it's not about random laptop memory, they just used the cheaper memory for the cheaper laptop
If Linus wasn't such an idiot, he wouldn't created so much confusion amongst his unfortunate viewers.
@@defeqel6537 Im not up on laptops but from what iv seen it's normal to change as ram comes in. Unless there sourcing from some one like Samsung or crucial it may not even be up to the laptop OEM what ram they get, lots of ram brands are pot luck on what ram brand is used on the sticks.
Iv even seen people say some laptops change from dual to single sticks and back during there life cycle.
Never mind during a parts shortage where it may not even be an option to stick to a single ram kind.
(and that ignores that most OEM's will at least dual source parts)
@@logmeindog he doesn't (usually) write videos, so blame one of his writers.
oh shit, buildzoid is evolving
nah I call this surrendering to the normies. Or a low effort view grab which ever you prefer.
@@ActuallyHardcoreOverclocking well, whatever you want to call it, I appreciate that you've put out a video explaining memory ranks that I can point people to without worrying that they won't understand.
@@GimpyChinaman YOU EVIDENTLY DIDN'T GET WHAT THE VIDEO IS ABOUT BECAUSE IT'S NOT ABOUT RANKS.
I AM SURROUNDED BY IDIOTS
@@ActuallyHardcoreOverclocking Triggered
@@ActuallyHardcoreOverclocking Maybe this person just got a semi-accidental answer to "why 4 ic, 8 ic and 16 ic sticks can all be single rank?".
Every other youtuber: "Oh man...I hope I can manage to squeeze at least 10 minutes out of the footage"
Buildzoid: "Oh man...I hope I can manage to keep focused and don't babble for 1.30 hours for this footage"
This was much easier to follow, even passively on the subway. Thanks for the update!
I think the biggest bullet point about RAM in this context is that most "normies" don't know that the bus width of DDR is 64-bit. This is then split into the above. 64-bit split into x4 is 16 modules, x8 is split into 8 modules, and x16 is split into 4 modules.
EDIT: First of all, thank you for the heart Buildzoid.
Since this is going to likely be a featured comment for a while, I would like to point out that what is being talked about in this video has nothing to do with memory ranks. The LTT video wasn't about ranks. Buildzoid's response to that video wasn't about ranks. This is about the SDRAM ICs on the DIMM, and the available 64-bit bus being split between all the modules on the DIMMs; see my original post for that summary.
You lost the heart by editing, been there, done that.
NAND is the nonvolatile storage chips in SSDs. SDRAM is what's used in DIMMs.
@@ActuallyHardcoreOverclocking for now at least. There's alot of people working to make RAM persistent. Game changer, no longer having to rewrite to each address row every y seconds depending the weakest capacitance in the row.
@@ActuallyHardcoreOverclocking Fixed.
Also 3d xpoint be eyeing DIMMs like:
Great series. After following them one has much better understanding of DDR memory, how it works and it's timing.
TL;DR: 64-bit RAM addresses are broken into x16, x8, or x4 bit banks, where each bank has a cost to access, write, read, and maintain data. The cost of performing a subsequent operation in the same bank is more than doing it in a different bank, thus the more banks, the better. As a result, x8 is better than x16 RAM configurations.
NL;PR: Random-access memory performance is directly proportional to the available slots to operate at once: dual-channel, x8 bit banks, dual-rank (2R) is the best consumer option there is so far; and to how often those slots can be operated on: higher frequency (>hz), lower latency class (
Thanks
Found it weird, that you had to make a new video explaining. Really think you did an awesome job in the other video. I'm really unfamiliar with the topic and understood every thing you were trying to explain, clearly.
This was incredibly helpful. Thanks so much!
BZ has surrender to the normie army with all the questions and chaos that Linus made with this video.
and I would say it's a great thing. That's how everyone becomes more educated on the topic i.e. how manufacturers are screwing their customers without letting them know explicitly
It should be tradition that "fast" words in the title are reserved for videos at least 40 minutes long.
You got through the whole video and even mentioned bus loading without having to explain ranks again, congrats :P
Great video, thank you BZ!
Really enlightening, much more than research papers.
Thank you for a really useful video, any chance you could do a follow up on how i could identify on desktop ram if a kit is x8 or x16 as i cant see an easy way to identify this before purchase
Good video. Thank you for sharing
This video was a lot better buildzoid. Abit of rambling but that’s just your as a person but u kept it on topic really will for one of your videos
What happens if you mix x8 and x16 sticks? does the x8 fall down to x16 access speed?
yes it falls down
@@claritoresdiano1021 Can you help me out pls ?
I have soldered memory of 4gb 2133mhz, 1Rx16 on my laptop. Should I add 4gb of x8 memory to it or 8gb of x16? (I cannot find 4gb of x16 in market)
@@akhilsingh9884 Open command prompt (Run As Administrator) then type:
wmic memphysical get MaxCapacity, MemoryDevices
to see the Maximum Capacity for each slot and the number of slots you have on your motherboard.
You could afford 8Gb x16 just go on that you get more headroom.
(4GB onboard) + (8GB Dedicated slot) = 10GB.
Don't bother about 4Gb x8 because your system can not maximize it.
Thanks, I understood it better this time :)
The big question is how do you know in advance whether its x8 or x16. I am looking at desktop RAMs and none of the big manfacturers (Corsair, G.Skill, Crucial) mention this on their website
See BuildZoid's other video. tl;dr: E-bay Corsair memory kits and look at the version in the listing's picture. You can look up what version corresponds to what memory module. The same exact SKU/product varies widely, so that's the only way. Alternately, you can try to find OEM or server parts, since they list plenty of extra details, but they're probably either going to be slow or expensive.
Thanks for the video. I never noticed this.
Thanks for the additional info.
Can you please explain mode registers ,write and read leveling ?
so i was wondering , does installing and x8 module with a x16 module cause any performance impact ?
Great very good explanation
Please explain the usage and what registers are in x4. Loved the video!
I don't care about server hardware.
Each pin on a memory IC has a tiny bit of capacitance (a parasitic property of the FET gates) and every time you add another memory chip to the bus you increase the total capacitance on the memory bus lines. The extra capacitance causes additional electrical "load" on the line, leading to slower transition times between defined voltage states (0 and 1). The longer the transition time, the less time the bus lines spend at a defined "1" or "0" state, making it more prone to errors. This is usually measured as an eye diagram. If you add a ton of memory ICs, you eventually end up way out of spec and the system can't function reliably.
In an unregistered DIMM (UDIMM) you've got a bunch of memory chips and they all share address lines, command lines, and (to some extent) data lines. All of these ICs have some parasitic capacitance that contributes to line load.
Registered DIMMs (RDIMMs) get around this by adding a buffer chip (a "register") between the memory ICs and the memory bus. Instead of connecting all the command and address lines from all of the chips to the memory bus, they instead connect them to a single buffer chip, and then connect that to the memory bus. The data lines remain connected directly from the memory ICs to the memory bus. The buffer distributes the address and command signals to the chips, which is easier for the buffer chip than the memory controller because it is physically close to the memory ICs, meaning there's less impedance in the signal path, and it can also have separate IO lines for groups of memory ICs, decreasing the load on each driver in the buffer. This approach heavily reduces the amount of electrical load on the command and address lines, allowing you to add more memory ICs in total, for better overall memory density. There's a latency penalty due to the added buffer, and the modules use a little bit more power too.
Load reduced DIMMs (LDRIMMs) go one step further and buffer the data lines too. This is more expensive since the buffer needs to handle many more signals, but it vastly reduces the bus load for all of the high speed signals, including data. This allows you to have far more memory ICs per module (technically it's more ranks, but explaining ranks is... ugh) and therefore huge memory densities on the standard DDRx bus. The penalty, again, is latency and power. Plus extra cost, of course, but that's usually not as much of a consideration when you're looking at 128GB DIMMs.
A rule of thumb is that UDIMMs use low density 8/16-bit chips, RDIMMs support high density 4-bit chips but don't support a lot of ranks because the data lines aren't buffered, and LRDIMMs support 4-bit chips AND high ranks (up to 8 iirc) because they buffer the data lines as well as the command & address lines.
(aside: there's also FBDIMMs but they're super frickin' weird and afaik nobody uses them any more. they use a serial bus to lower the number of lines that need to be driven, then expand that back to a parallel DDR bus on the DIMM using something similar to an LRDIMM register buffer but waaaaay more complicated. a lot of that tech ended up being vapourware.)
@@ericrose3978 Yeah, Skulltrail and some SPARC stuff used it, but it was very much a niche thing.
@@ericrose3978 Please do! That stuff is pretty rare now, and should be preserved as much as possible.
Nice explanation.
So, the fewer the number of bank groups is, the more latency there will be for read/write cycle.
Great format.
When I press play is it Steve from GN with a summary???
this was so much easier to understand, thank you!
Hello,
Can you make a vid with a am4 + a 5000 series cpu and a b-die oc to 4000+ cl 14 running fclk 2000+ ?
This is a fucking win. good job man.
If there is any interest, could you do a partpicker walkthrough of available 2x32GB kits? I was looking through your previous videos and didn't find anything like this specifically.
Crucial ballistix (rev b) is the answer
@@rayanzoghlami6682 Thanks! I ordered BL2K32G36C16U4B from amazon. I really hope that is the right ones.
@@sunefred yep, that’s the one 👍
how do i know if my ddr4 is 4x,8x or 16x?
This video isn't really a simplified version of the other, it's a complement.
The video explains why the realities of manufacturing dictate why x16s have a performance penalty while the other has discussion on the qualities of different dies and whether this information is made available to the consumer.
thought it was fine the first time, but ok ill watch this too, guud job
I feel like Buildzoid is more appropriate holder of the title "tech Jesus"
Can you fix cc please? Hearing technical number in noisy loud environment is kinda... Anyway, thanks for vid.
Thanks. That made sense. It could also be helpful if you started with a photo of an RAM entire module, and then break that down into ranks, and then into the individual chips, and then into the banks in the chips. Perhaps with examples of 1Rx8 vs 2Rx8 (dual rank ryzen memory debacle) vs 1Rx16 (stupid laptop manufacturer debacle)
Anyone know how to differentiate between x8 and x16 when buying from amazon?
I just bought some Hynix that are x8 that is the same as the one that came in my laptop, i saw that Samsung also tells you in the label
@@dawicked2k8 Thanks!
So if x16 types of IC's are the worse type of IC's to use how can companies sell it as performance Desktop ram and use it for XMP?
I have X16 ram in first slot but I ordered X8 for another slot will they work finely together?
X4 vs x8
What's better for legion 5 pro?
Please am waiting for your advice
Question, can I identify the x8 or x16 by reading specs on CPU-Z?
Aida64 and HWiNFO MEM specs screenshots would make this video perfect.
Huh, I thought the original video was perfectly understandable. How I interpreted it: if you want a faster HDD array, add drives so you can have more seeks going on at once; if you want a faster RAM array, add bank groups so you can have more selects going on at once...
what is memory density ?
Can the performance penalty be mitigated by having more ranks or more than 4 chips per memory stick?
sort of, yes, because then the memory controller would have 2 ranks with 2 bank groups each to use instead of 1 rank with 2 bank groups. But for the applications where this could matter (laptops etc) you won't have the opportunity to do that because you're limited to 2 SODIMMs anyway and they'll be single rank. For desktops you'd much rather have 2 ranks with 4 bank groups each
if they actually made x16 chips with 4 bank groups instead of 2, then they could make dual rank SODIMMs (same number of chips as single rank with x8 chips) with 4 bank groups which would be pretty awesome for laptops
Subscribed
so what you're saying is if i take the bus to the bank it's slower than taking a car?
So glad you made this video. 🙏 Can x4 theoretically give a performance boost if the density is kept low? I have been trying to understand the RAM options for the Mac Pro 7,1 for a long time! Because nerdy.
x4 = 16 chip
x8 = 8 chip
x16 = 4 chip
Theoretically is yes but the fact is impossible. SODIMM is small no headroom for 16 chips (x4), that's why just x8 or x16
is this memory good?
PSD516G4800K, 16GB (2 x 8GB) Kit, 4800MHz CL 40, 1.1V
I do have x4 laptop sticks because the old PoS Core 2 CPU in an old laptop I have can't handle x8 4GB DIMMs, but is fine with x4 4GB DIMMs, so it was the only way to get 8GB in that machine. The only reason this travesty even exists is because some old Macs had the same issue and of course HW manufactureres will make any garbage for Macs (it's actually sold as "Mac memory").
Give us some ddr5 thoughts !
Explained it fine to me mate...
which one is better 16gb 2rx16 or 16gb 1rx8?
Well, and a typical Bdie 16 GB dual rank dimm is what exactly?
Thanks BZ
I don't think the problem is tccd. I think the issue is the fewer number of banks increases the probability of needing to precharge the current row and activate a different one. The tccd seems to me like it would be way easier for the controller to schedule around.
Can you help me out pls ?
I have soldered memory of 4gb 2133mhz, 1Rx16 on my laptop. Should I add 4gb of x8 memory to it or 8gb of x16? (I cannot find 4gb of x16 in market)
Clear as mud!
Is there a TLDR short conclusion to watch that?
Buildzoid, what is your degree in school?
failure
@@ActuallyHardcoreOverclocking ah! we went to the same school.. although we moved on and now your a big guy on youtube and im a person watching your videos.
Hello, Have a look at the EDC bug on the 3000 ryzen. I have my 3900x boosting to 4.7ghz on half of the cores crazy stuff. Just use pbo, put EDC 1 clock to +500mhz escalar to 10* and then I did a bit of undervolt -0375 and push the bus clock to a point before 101,00.
What is this autogenerated Vietnamese subtitles? Why not english for non native speakers? it is night and day difference when there is proper transcript and I put on 0.75 speed and replay as needed.
Ok so if i get this correctly, if some company decides to make x16 chips with 4 banks, they'd be (approx) the same as x8's right? (because they have the same amount of bank groups)
And is it possible to have more than 4 bank groups or is that outside of the spec?
(idk if this got answered in the long vid, that one did go a lil over my head)
Did you learn the answer?😅
Although I am late with this comment, I would like to thank you for the in-depth explanation. It makes things clearer, especially that a branded memory stick is almost always x8 but non-branded may very well be x16. But I do want to ask how you know what type you have in your laptop. Do you really have to open it up and read the sticker? Is there no software like CPU-Z or HWInfo64 that can tell you? I may be a n00b, but I can't find this info from these two programs.
Youre gonna have to open it up to match the one you already have, thats what i did
Can you help me out pls ?
I have soldered memory of 4gb 2133mhz, 1Rx16 on my laptop. Should I add 4gb of x8 memory to it or 8gb of x16? (I cannot find 4gb of x16 in market)
Akhil Singh in my case I dont see too much of a difference for everyday things that I do, mainly browse the internet and the only game I play is MW2 which is an old game which is not demanding, its always nice to have more RAM for when you do need it, my RAM was not soldier so that was a plus. I just match the one i ready had.In my case Ive read that AMD Ryzen CPUs are more sensitive to the right RAM.I guess with Intel you can mix and match different RAM modules. Having said that tho if you mix x8 with x16 you wont see the speed inmprovements of x8, Im not a expert but I know for my laptop I got the best RAM I could get, another thing my RYZEN cpu cannot use 3200mhz, I was gonna buy that speed because it was the fastest but I read that 2600mhz was the maximum, so if i had bought that speed I would have just wasted money, I would advice watching more RAM videos here on TH-cam, about your particular laptop and CPU, theres lots of giod videos here
thank you
TH-cam: This guy has a terrible accent don't you think?
Caption algorithm: Sure, somewhat asian style, I heard Micron and Samsung.
TH-cam: Let's assume he's vietnamese then.
Caption algorithm: Agreed. Starting auto-translation... "that's what going on with a, with-aaa bus width... ghi chép tât phen nuôi work best" Matching voice 99%.
first
Second? Love your channel. Great job.
Can you drop the link to that Micron document? Thx
Manufacturers can solve this problem with 3D layered chips or stacking more die in a single package. I.E. High capacity are typically stacked in an 8 die per package format. The would need to use a higher density BGA.
cooling it will be an issue though, gpus and cpus probably can work with direct airflow.. not sure about ram sticks ?
This might be the first time I actually consider recommending a buildzoid video to normies.... well done!
a video that is only 10 min?? lmao
So 2Rx8 is best?
yes
x4 vs x8 which is better for gaming ?
x8, x4 is for servers
what the differences between 2 rank and x16 ( physical ) ?
One rank is 64 bits of memory interface. So x4, x8 and x16 only relates to rank in the sense that it dictates how many physical modules are needed to reach the 64 bits needed per rank.
Awesome video!! This is super new to me. How can I identify an x16 vs an x8 in desktop memory?
see on the memory stick, less chips means x16
So they could make a four bank x16 chip. But nobody does.
yes
Ok I had a good grasp on this after the first video. Now I’m confused 😂
Say you have a few filing cabinets at your office, but if you open them too quickly it jams, so you have to do it slowly. Now lets say that Corporate has come down and put a thing that auto-closes them after every page you take it/put in. It takes 4 seconds to open a cabinet, and 2 seconds for the auto-closer to work. So, if you only have a single cabinet, you would have to wait 6 seconds between each page. However, if you had two cabinets, then you could try to spread the document between them, and only have to wait 4 seconds. Except, sometimes you end up needing two pages from the same cabinet. Well, with four cabinets, that's much less likely to happen. So, it's faster.
That's an analogy for what's happening. x16 is like having two larger cabinets instead of 4. They can each hold more pages, but you're much more likely to have to wait the full 6 seconds.
Incidentally, in this analogy, the cabinets are RAM. Your desk is cache, and corporate's main document repository across the country where documents have to be overnighted is your SSD.
Your comment about servers using x4 is a bit inconsistent.
An x4 chip would have, going by the 4Gbit x4 example, the same capacity as the 2Gbit x8.
You can see this on the datasheets: 262144 x 128 x 32 = 1,073,741,824 bits, is the same as 131072 x 128 x 64.
Not sure really why you'd use x4, maybe easier PCB layout for lower speed controllers. Or maybe they're cheaper to make for non-dense applications.
x4 chips require 16 chips per rank which is hard to drive so they only get used on RDIMM. RDIMMs are for severs because consumer chips don't even support those.
You have to have 64-bits worth of memory ICs in a rank. If you use 8-bit ICs, you can only have 8 of them comprising a single rank. That limits the total per-rank density to the maximum capacity of available memory ICs multiplied by 8. If you use 4-bit ICs, you can double that and get twice the density per rank. The problem is that all of those address and command lines are shared, so the load on the bus (largely due to parasitic capacitance on the memory IOs) is doubled when you do that. That's why you only see them in buffered memory (i.e. registered DIMMs) and not in consumer memory.
@@ActuallyHardcoreOverclocking Right - I get that. My comment was more around how they'd put more chips on an RDIMM to get higher memory capacity, but this wouldn't work like that. Each chip has the same capacity, just differing geometry. The only reason to do this that I can see is if you wanted a 32-bit bus instead of e.g. a 64-bit bus, but the same capacity. Maybe for an embedded processor that didn't have the pins for 64 bits.
Here's a 2Rx4 RDIMM: uk.crucial.com/memory/server-ddr4/mta36asf8g72pz-3g2b2
@@ActuallyHardcoreOverclocking I think you're missing out a bit here,my laptop has x4 memory 3200mhz 8gb each,if I go by your facts this ram shouldn't exist.
Thanx man! It was explanation on ranks and banks, I was just reading about ram and came across this term "absolute latency" calculated from ram speed (MHz) and its CAS latency (ns) on online ram latency calculators... and i was wondering is that a good enough indication of ram speed overall ?
So X4>X8>X16 in terms of memory speed? And also please explain why X4 laptop memory chips are not feasible?
The equation is * = . Where the chip width is x4, x8, x16, and the bus width is always 64 per stick of ram. So it's 64/ = . That means:
x4 = 16 chips
x8 = 8 chips
x16 = 4 chips
You can't physically fit that many chips on a SODIMM.
2Rx8 > 1Rx8 > 2Rx16 > 1Rx16
You said that x4 are density optimised and x16 are for small PCBs. From my understanding the goal in both cases is to cramp es much bit cells in a given space. Why is one ideal for server and the other ideal for Laptops?
If people didn't understand your first video....... well, you see where this is going. It's like teaching algebra to dogs, cute & entertaining but ultimately futile.
Amazing video, thank you Buildzoid! This makes me wonder about CPU Cache differences between L1 L2 L3 or even just more modern CPUs and older CPUs. It also makes me think about Nvidia’s various Graphics Card SKUs and how they have all these different memory bus widths, but it’s a set amount per 2 chips and that dictates capacity with the exceptions of higher capacity chips as found on the 3060 (non-Ti) and 3090.
The 3090 uses the same density chips as the 3080(Ti) it just uses twice as many chips
@@ActuallyHardcoreOverclocking Oh I assumed it was double the density kind of like the RTX 2080 Ti 11GB to RTX TITAN being +1 chip to get 24GB. I didn’t think the 3090 had space on the PCB... hmm double sided i guess? Anyways thanks for pointing that out.
1 *B* = 8 *b* (1 byte = 8 bits)
8 MB/s = 64 Mbps (8 megabytes/second = 64 megabits per second)
16 Gb = 2 GB
8 Gb = 1 GB
4 Gb = 0.5 GB
Decimal
10^3 B = 1,000 B = 1 kB (1 kilobyte)
10^6 B = 1,000,000 B = 1 MB (1 megabyte)
10^9 B = 1,000,000,000 B = 1 GB (1 gigabyte)
Binary
2^10 B = 1,024 B = 1 kB (or 1 KiB, i.e. 1 kibibyte...)
2^20 B = 1,048,576 B = 1 MB (or 1 MiB, i.e. 1 mebibyte...)
2^30 B = 1,073,741,824 B = 1 GB (or 1 GiB, i.e. 1 gibibyte...)
Any tips on upgrading memory on a laptop? It's a Tuf A15 with 5800H with a 3060 and 16g of ram.
I think you could make this video much simpler and much better. Basically you just waved your mouse around and didn't explain the diagrams at all. Maybe you expected me to pause it and find the differences myself. Then you say they're the same chip so I guess the difference is just in the start of diagram, I don't know. You didn't say look over here, there's X amount of banks or X amounts of data lines or whatever. You just talked a bunch just to say they're the same chip that can be configured differently. I'm not sure if you explained why the timing is different based on configuration used. It seems like you just said one configuration needs to wait more clock ticks to access same block or bank or whatever that you didn't explain. It would of been nice to know what the block sizes are and so I get an idea how often it will try to read from same block or sector or whatever you call it. Part of me wants to be nice and try to help you and part of me is like omg that's the worst explanation ever so you realize how bad it was so I guess I'll just say both. I suggest just draw 3 squares, put 4x 8x and 16x then inside draw the blocks all the same size, tell us size, maybe draw a controller block and say this addresses them different. I rather a rough drawing that is explained then a complex one that is not.
I bought a sodimm from adat 8g 3200mhz it says 8Gx4 on the ram sticker and the chips are only on one side of the ram if I have a xr ram is it much better than x8?
too much talk ... just say it please, which is faster which is better x4 or x8 or x16 ???
A sub 15 minute video? What sorcery is this?
So a DIMM with 8 chips is best. If it has 16 then there is more bus load. And you really don't want 4 because of the lower performance.
If you ever find a compact notebook with the RAM soldered down to the motherboard it may be x16 simply to make room. Linus did a video on x8 and x16 RAM in identical computers. one came with x8 and one came with x16. x8 is seemingly better for performance somehow. It really is ironic but you would think the higher the number the better it surely must be. If you have one of those people who keeps saying more is better and doesn't know much about computers just send him on a wild goose chase looking for x32 RAM (which does not exist hehehe!!). Tell him not to come back until he finds it
...As a follow up to my previous comment, my main concern is that you'll start spending more time explaining basic s*it and less time being cryptically hardcore...which is why i tune in.
Just keep being you and don't worry if the norms are keeping up. Google is a thing and learning is a skill that needs daily practice.
some one was telling me, that the width could be set after the chip was made, lol. BGA pin-out I guess is not "programmable" lol.
Let me condense this further for the gamers: *_Buy the x8_*
And For Mac users let me tell you
M1 is using x64 sticks 🥲
That sucks
Please don't keep scrolling the screen all the time.
X-16 +- 8 - Calculus rule - 4 is, far too small.